Building an Agentic Platform - Technical Evolution of Metadata Extraction at Box

Discover the technical journey of transforming metadata extraction at Box through a comprehensive 19-minute conference talk that chronicles the evolution from traditional LLM approaches to agentic architecture. Learn about Box's enterprise content platform challenges with unstructured data and follow the detailed progression from their initial 2023 AI deployment through the limitations of pre-generative AI extraction methods. Examine the first-generation LLM-based extraction system and understand the critical challenges that emerged with pure LLM approaches, leading to architectural despair and the urgent need for redesign. Explore the introduction of agentic architecture as a solution, including the AI agent reasoning framework and specific agentic routines developed for data extraction. Analyze the significant advantages unlocked by this architectural shift and absorb the key lesson about building agentic systems from the ground up rather than retrofitting existing solutions. Gain insights into Box's approach to fine-tuning and model support strategies, presented by Ben Kus, CTO of Box, who brings extensive experience from leadership roles at IBM, BigFix, and as co-founder of Subspace before its acquisition by Box.